DeepSeek V4 Flash 0731 (TensorX)

Eden AI's blog (August 24, 2026) directly references DeepSeek V4 Flash 0731 in its benchmark table alongside the Vision Exp variant, reporting Terminal Bench 2.1 at 82.7 and ApexBench Pass@1 at 26.2 for the 0731 checkpoint. The post compares the Vision variant against Claude Opus 4.8, noting DeepSeek wins 3 of 11 self- The piece explicitly confirms Eden AI has DeepSeek V4 Flash 0731 in scope, but is transparent about using DeepSeek's own Harness Minimal Mode evaluation with no independent verification, and notes Opus 4.8 pricing as unverified. It is a vendor-authored comparison post rather than an independent benchmark or release ann

Artificial Analysis (July 31, 2026) independently scored DeepSeek V4 Flash 0731 at 50 on the Intelligence Index, a 10-point jump over the prior V4 Flash and within 1 point of GPT-5.6 Luna (max). The model shares identical 284B/13B-active architecture and 1M context with its predecessor, lands on AA's Pareto frontier fo Terminal-Bench 2.1 rises 17 points to 79% and τ³-Bench Banking 8 points to 31%. AA-Omniscience Index improved to -16 (from -23) driven by a 12-point hallucination rate reduction. DeepSeek's 98% cache-hit discount on its first-party API makes 0731 60% cheaper per task than GPT-5.6 Luna (max). Full weights were expected

Opper's gateway page for DeepSeek V4 Flash (0731) confirms a July 31, 2026 release date for the retrained checkpoint that took the V4-Flash API into public beta, with an unchanged 284-billion-parameter sparse MoE architecture and 13 billion active parameters per token. The model retains a 1M token context window, 393K The page lists 10 provider routes through the Opper gateway including TensorX EU at $0.25 input / $0.30 output per million tokens, matching the Eden AI routing path. It confirms three reasoning modes (non-think, think-high, think-max) and XML-based tool-call format, positioning the 0731 build as a drop-in successor to

DeepSeek's official API changelog documents the August 21, 2026 release of DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal vision model accessible via model='deepseek-v4-flash-vision-exp', with self-reported benchmarks including Terminal Bench 2.1 at 83.9, NL2Repo at 57.7, and ZeroBench Pass@5 at 35.0. The cha While the dated entries cover V4-Pro and Vision-Exp rather than the 0731 checkpoint directly, the changelog confirms that V4-Flash supports the three-mode thinking control and that the model family is actively evolving with native Responses API support and Codex integration. This is the highest-quality primary source f

KissAPI's blog index includes an August 24, 2026 entry titled 'DeepSeek V4 Flash Vision Exp API Guide (2026): Vision, Pricing, and Routing' which covers image and screenshot understanding added to DeepSeek V4 Flash, along with API usage, pricing, routing, and implementation patterns for multimodal apps. The listing con Rejection note: The KissAPI listing only provides a title fragment and date for the Vision Exp variant guide; it does not contain substantive technical detail about the V4 Flash 0731 checkpoint itself and is not useful as a standalone reference.